Research On Unsupervised Person Re-Idenfitication Based On Deep Asymmetric Metric Learning

In recent years,security industry and emerging the comprehensive integration of information technology is accelerating.Cross-view joint video analysis has become an important tool to maintain urban public safety and social stability.As a key content of joint video analysis,person re-identification has shown great potential in scenes such as rapid identification and tracking of criminal suspects,and tracking of the activity trajectory of epidemic infected persons.However,the dynamic change of cross-domain pedestrian data makes manual labeling of large-scale pedestrian samples impossible,and the dynamic change of imaging environment makes the distribution difference of cross-domain pedestrian features increase,and it is difficult to maintain the spatial distribution consistency.Therefore,it is of great theoretical significance and practical value to learn from the idea of unsupervised metric learning to carry out the research on feature extraction strategies and metric learning methods in cross-domain person re-identification.The main innovations of this paper are as follows:(1)Traditional unsupervised asymmetric metric learning can overcome pedestrian feature differences caused by different camera angles,but it is difficult to overcome pedestrian feature differences caused by scene changes,light changes and other factors.Therefore,this paper proposes an unsupervised person re-identification method based on distribution-constrained asymmetric metric learning.Based on the unsupervised asymmetric metric learning objective function,a new linear asymmetric metric learning objective function is defined by introducing distribution constraints,and the objective function is solved by gradient descent method.Finally,the objective function solution is transformed into a generalized eigenvalue problem.The proposed method not only maintains the advantage of traditional asymmetric metric learning that can overcome the view interference,but also achieves the purpose of overcoming the inconsistency of pedestrian feature distribution caused by scene changes and other factors,and effectively improves the effect of unsupervised person re-identification.(2)To solve the problem that pedestrian features are not linearly separable,an unsupervised person re-identification method based on kernel distribution constrained asymmetric metric learning was proposed.All the features of the original feature space are projected to the high-dimensional feature space by a nonlinear mapping.Due to the large dimension of the high-dimensional feature space,it is easy to generate dimensional disaster and the amount of computation is large.The "kernel technique" is introduced to define the objective function of the asymmetric metric learning with kernel distribution constraints.Finally,the gradient descent method is used to transform the objective function into a generalized eigenvalue problem.This method not only overcomes the problem of inconsistent distribution of pedestrian features caused by factors such as scene changes,but also solves the problem of nonlinear features in pedestrian features,which effectively improves the effect of unsupervised person re-identification.(3)Traditional unsupervised asymmetric metric learning methods do not jointly optimize feature representation and metric learning.To solve this problem,an unsupervised person reidentification method based on deep asymmetric metric learning was proposed.The asymmetric metric learning was embedded into the Feature extraction Network(ResNet)for end-to-end joint learning,and a new loss function was constructed.This method can not only alleviate the bias caused by different views,but also help to mine potential discriminative features.